Skip to content

THIRD_PARTY_NOTICES: the tensor-parallel fabric's RoCE attribution beside the CUDA gather's (carried from #636) #1128

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
jayleaton wants to merge 1 commit into from
Closed

THIRD_PARTY_NOTICES: the tensor-parallel fabric's RoCE attribution beside the CUDA gather's (carried from #636) #1128

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions THIRD_PARTY_NOTICES.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,6 +189,10 @@ the one-shot all-gather protocol of [b12x](https://github.com/local-inference-la
sequence parity and a timeout that fails the transport. Copyright 2026 Luke Alonso and the b12x contributors,
Apache License 2.0 (see [the license text](LICENSES/Apache-2.0.txt)); b12x ships no NOTICE file.

The Zig tensor-parallel fabric's RoCE path (`zig/src/fabric/tp/mailbox.cu`, `roce.zig`, `roce_verbs.zig`) adapts
the same one-shot all-gather protocol of b12x's RoCEnante (`b12x/comm/roce/`, PRs #295 and #315) for any family that
runs tensor parallelism over it, under the same copyright and Apache License 2.0 terms.

## Vendored code and weights

In the Python engine, frozen at 0.6.6, `src/tensorfold/drafters/vendor/z_lab_dflash/model_mlx.py` is the unmodified
Expand Down Expand Up @@ -342,8 +346,6 @@ common EXL3 operators and their per-slot epilogues; tile decoding is shared.
The ExLlamaV3 MIT attribution, Copyright (c) 2025 Turboderp, applies as described above.
No external reference, captured cubin, model shard or CUDA compiler is bundled by this patch.

The tensor-parallel fabric also uses the RoCE design described above (`zig/src/fabric/tp/mailbox.cu`, `roce.zig`, `roce_verbs.zig`).

## Shared EXL3 resolver setup

The shared EXL3 resolver split and CUDA lane option forwarding retain work from PR #728 by @ajensenwaud,
Expand Down
Loading